Add infrastructure for changing acquisition function optimizers #482
Conversation
The BotorchStrategy is expanded to modularize the acquisition function optimizer. Every acquisition function optimizer has to derive from a newly created base class. The current Botorch-based optimizer is moved into a separate class. This includes a breaking change in the data model: A new data model for acquisition function optimizers is created. The Botorch-optimizer params stored in the BotorchStrategy are moved to the new BotorchAcqfOptimizer.
|
Thanks @sleweke-bayer! I will have a detailed look on this in the time between the years! |
There was a problem hiding this comment.
Hi @sleweke-bayer,
thank you very much. I do not had a detailed look but focused more on the architecture and had especially one comment there. What do you think?
Best and a happy new year,
Johannes
| return cls(data_model=data_model) | ||
|
|
||
| @abstractmethod | ||
| def optimize( |
There was a problem hiding this comment.
From my understanding the signature of the optimize method should be a bit simpler and maybe also more torch agnostic. From my perspective, we only need the domain, the bounds, and the acqf(s), the rest can be generated on the fly from domain and is also the optimizer specific, or?
I mean that for example a GA wants to have the constraints in a different format then botorch wants them. In principal, also the bounds could be generated on the fly out of the domain, but this would create problems with the local and global bounds, as this is controlled on a different level. Of course, one could also think about putting the LSR logic into the generic optimizer class, then we would only need the acqfs and the domain and maybe also the input processing specs.
What do you think about making it more agnostic here?
The
BotorchStrategyis expanded to modularize the acquisition function optimizer. Every acquisition function optimizer has to derive from the newly created base classAcquisitionFunctionOptimizer. The current Botorch-based optimizer is moved into the separate classBotorchAcqfOptimizer.This PR includes a breaking change in the data model: A new data model for acquisition function optimizers is created. The Botorch-optimizer params stored in the
BotorchStrategyare moved to the newBotorchAcqfOptimizer.Due to the modularization, we can later add different optimizers for the acquisition function (e.g., genetic algorithm).